By: Adrian (a.delete@this.acm.org), May 23, 2022 12:25 am
Room: Moderated Discussions
Jörn Engel (joern.delete@this.purestorage.com) on May 22, 2022 11:51 pm wrote:
>
> I'm with Linus on this one. Having fast and slow cores is mostly fine. But having cores with fundamentally
> different behavior is too much pain to be worth it, with embedded systems as a possible exception.
ARM has designed the much smaller core Cortex-A510 to also support SVE2, like the medium-size core Cortex-A710/Neoverse N2 and the large core Cortex-X2, the latter 2 cores having complexities comparable with the Gracemont and Golden Cove cores of Alder Lake.
This proves that the heterogeneous ISA that is traditional for the Atom/Core CPU series is only due to Intel being either stupid or mean, and not due to any technical necessity.
So I agree that it is not the programmers who should work to implement a way to cope with this misfeature of the Intel CPUs, but it is Intel who should abandon their evil ways and follow completely the ARM Big.Little way, including a uniform ISA, if they have started to go in that direction.
The uniform ISA must not be achieved by disabling features on the small cores, as that has the consequences that we have seen with AVX-512, which, much more than a decade after its first announcement, is not much closer of being a widespread target for software development.
>
> I'm with Linus on this one. Having fast and slow cores is mostly fine. But having cores with fundamentally
> different behavior is too much pain to be worth it, with embedded systems as a possible exception.
ARM has designed the much smaller core Cortex-A510 to also support SVE2, like the medium-size core Cortex-A710/Neoverse N2 and the large core Cortex-X2, the latter 2 cores having complexities comparable with the Gracemont and Golden Cove cores of Alder Lake.
This proves that the heterogeneous ISA that is traditional for the Atom/Core CPU series is only due to Intel being either stupid or mean, and not due to any technical necessity.
So I agree that it is not the programmers who should work to implement a way to cope with this misfeature of the Intel CPUs, but it is Intel who should abandon their evil ways and follow completely the ARM Big.Little way, including a uniform ISA, if they have started to go in that direction.
The uniform ISA must not be achieved by disabling features on the small cores, as that has the consequences that we have seen with AVX-512, which, much more than a decade after its first announcement, is not much closer of being a widespread target for software development.
